This March, Sofya announces that its research has been accepted at the traditional spring symposium of the Association for the Advancement of Artificial Intelligence (AAAI), hosted at the Clinical Foundation Models at Stanford University. The Clinical LLM (soon to be open-source) surpassed GPT3.5 and Mixtral7x8b and achieved comparability with Google-Gemini Pro in clinical data structuring tasks using health interoperability standards like HL7 FHIR FOUNDATION entities.
The objective of the research is to facilitate better healthcare planning and efficient, personalized patient care through comprehensive and organized clinical documentation. This is because the intense workload has led to a concerning growing trend: 50.4% of physicians reported burnout (Ortega et al. 2023), and nearly 800 thousand patients annually in the United States (US) are harmed by diagnostic errors. Many of them were associated with cognitive errors, according to a recent study by John Hopkins (Newman-Toker et al. 2024).
"Physicians typically deal with high density and noise of narrative data. Providing them with this level of structure and summary helps alleviate cognitive pressure and the dramatic situation of overload in healthcare delivery that healthcare professionals experience today, as well as a leap in the quality and personalization of patient care," says Sofya's CEO, Igor Couto.
Sofya's research addresses a practical and crucial need in clinical workflows, potentially paving the way for deeper insights and intelligence on patient care in the future. In the scope of this work, Sofya employed a basic open-source multilingual model to solve the fundamental challenge of patient data structuring.
To ensure robust and universally applicable solutions, the model was initially aligned and optimized for three distinct subtasks, each focused on a different aspect of patient information and interaction with the healthcare system: Patient Clinical Summary (FHIR IPS), Clinical Impression, and medical encounter structuring.
To build a Clinical AI model, the approach focused not only on building LLM capabilities but also on enabling global digital health standards such as the International Patient Summary (IPS) and the HL7 Fast Health Interoperability Resources (FHIR).
"In the work, in addition to domain expertise for clinical data problems and healthcare workflows, we faced challenges such as context size, dataset expansion, and rapid and economical experimentation. We also demonstrated the results with an automated blind method of pairwise with control groups, allowing teams to quickly measure model evolution", Couto points out.
No final de março, a Sofya apresentará o trabalho na conferência Associação para o Avanço da Inteligência Artificial (AAAI), no Clinical Foundation Models da Universidade de Stanford, bem como serão divulgados os pesos dos modelos e conjuntos de dados de treinamento/avaliação.
At the end of March, Sofya will present the work at the Association for the Advancement of Artificial Intelligence (AAAI) conference, at Stanford University's Clinical Foundation Models, as well as release the model weights and training/evaluation datasets.